Image Colorization Based on Two-stage Convolutional Neural Networks

• Main Authors: CHEN, ZHENG-DA, 陳政達
• Other Authors: KUO, TIEN-YING
• Format: Others
• Language: zh-TW
• Published: 2019
• Online Access: http://ndltd.ncl.edu.tw/handle/384s37

碩士 === 國立臺北科技大學 === 電機工程研究所 === 107 === In human visual habits, color images can provide more information than grayscale images. If we can color grayscale images, it not only increases the visibility of the images, but also provides more image information. At present, most of the papers mainly use convolutional neural networks. Although they have good coloring effects, the complexity of the network architecture is too high. Therefore, this paper proposes two-stage networks architecture to reduce the complexity. This work proposes an image colorization framework based on a two-stage convolutional neural network architecture. First, we convert the training images to the CIEL*a*b* color space and enhance the saturation of training images in the pre-processing. In the first stage, we use a convolutional neural network to generate a low-resolution chromatic map, which is then scaled up to two different resolutions using an image pyramid alike architecture. The two different resolution images are input into the refined network of the second stage at the same time and trained by shared weight, to obtain feature maps with different resolutions. Both the feature maps are enlarged to the original input image resolution using the transposed convolution network. Finally, the enlarged feature maps are combined and output a finer color image.